International Business Machines Corporation

Faraydon O Karim

Christopher Hans Olson

Michael Thomas Dibrino

Tan V Chu

Presented is a &ldquo;high-order&rdquo; Leading Zeros Anticipator or LZA circuit and specifically a five-input LZA. The prior-art two-input LZA circuit is part of almost all high-performance floating-point units or FPUs. The advantages of a high-order LZA (such as five-input) is that the LZA function may be started and finished sooner in the floating point pipeline, and therefore allows more time for other functions in the pipeline. Therefore, a high-order LZA, such as five-input LZA, may be faster than the prior art two-input LZA designs. Thus, speeding up the LZA function in a floating point pipeline may significantly increase the speed in which the overall floating-point unit may operate as compared to the prior-art two input LZA designs and may additionally inspire new floating-point michroarchitectures which may yield further performance gains.

Floating point pipeline with a leading zeros anticipator circuit

An apparatus for generating an end-around carry to an end-around carry adder in a floating-point pipeline within a computer system is disclosed. The apparatus for generating an end-around carry includes a shift-comparison logic circuit, a sign-comparison circuit, and a logic gate. The shift-comparison logic circuit produces a shift-count signal and the sign-comparison logic circuit produces an effective operation signal. Coupled to the shift-comparison logic circuit and the sign-comparison logic circuit, the logic gate combines the shift-count signal and the effective operation signal with a carry-out signal generated by an end-around carry adder to provide an end-around carry signal for the end-around carry adder.

Method and apparatus for generating an end-around carry in a

A super scalar computer architecture and method of operation for executing instructions out-of-order while managing for data dependencies, data anti-dependencies, and integrity of sequentiality for precise interrupts, restarts and branch deletions. Multiple registers and tables are used to rename and recycle source and destination addresses referenced to a general purpose register. Access to destination data in the general purpose register is locked until the instruction associated with the data is fully executed. Renaming of both the source and destination registers avoids anti-dependency problems while integrity of sequentiality is maintained by ordered retirement of instruction results consistent with the order of the input instructions. The system and method operate with multiple input instructions and multiple execution units. The control words generated by the renaming of the source and destination registers differ insignificantly from the original instructions, obviating the practice of adding status and sequence information to processor control words.

Super scalar computer architecture using remand and recycled general

A method and apparatus for performing normalization of floating point numbers using a much smaller width register than would normally be required for the data operands which can be processed. As the registers are smaller, the number of circuits required to achieve the normalization is reduced, resulting in a decrease in the chip area required to perform such operation. The normalization circuitry was streamlined to efficiently operate on the more prevalent type of data being presented to the floating point unit. Data types and/or operations which statistically occur less frequently require multiple cycles of the normalization function. It was found that for the more prevalent data types and/or operations, the width of the registers required was substantially less than the width required for the less frequent data types and/or operations. Instead of expanding the register width to accommodate these lesser occurrences, the data is broken into smaller portions and normalized using successive cycles of the normalization circuitry. Thus, by sacrificing speed for the lesser occurring events, a significant savings was realized in the number of circuits required to implement normalization. As the slower speed operations occur infrequently, the overall performance of the normalization function is minimally impacted. Thus, considerable savings in integrated circuit real estate is achieved with minimal impact to the overall throughput of the system.

Method and apparatus for floating point normalization

An architecture and method relating to a floating point operation which performs the mathematical computation of A*B+C. The multiplication is accomplished in two or more stages, each stage involving corresponding sets of partial products and concurrently accomplished incremental summations. A pipelined architecture provides for the summation of the least significant bits of an intermediate product with operand C at a stage preceding entry into a full adder. Thereby, a significant portion of the full adder can be replaced by a simpler and smaller incrementer circuit. Partitioning of the multiplication operation into two or more partial product operations proportionally reduces the size of the multiplier required. Pipelining and concurrence execution of multiplication and addition operation in the multiplier provides in two cycles the results of the mathematical operation A*B+C while using a full adder of three-quarters normal size.

Floating point arithmetic unit with size efficient pipelined

An arithmetic circuit for performing an integer power operation having a first register that is initially stored with integer power data, a second register stored with base data and a multiplier for multiplying the second register contents with a third register. A control circuit is provided for iteratively shifting the first register and for each iteration (1) selectively multiplying, according to data shifted from said first register, the second register data by the third register data and storing the results in the third register, and (2) multiplying the second register data by itself and storing the results back into the second register. These iterative steps are continued until the remaining data in the first register indicates that the arithmetic operation is complete, whereupon a completion signal is provided.

Circuit and method for performing integer power operations

A system for detecting multiple errors that may occur during transfer of data and for correcting up to two of these errors simultaneously. The system has a component for calculating a number of check bits associated with the data word. Also provided is a component for grouping all data bits into base groups and multiple groups, the sum of the number of base groups and multiple groups being equal to the number of check bits. Up to two weights are assigned for each data bit. The system distributes the data bits among the groups according to the weights assigned thereto. Also provided is a component for generating a check bit for each of the groups and for padding the data word with the check bits to form an appended data word. A generator creates a predetermined number of syndrome bits, the number being the number of check bits. Finally, a decoder is provided for decoding the syndrome bits to identify the erroneous bits in the data word.

Growing community of inventors

Round Rock, TX, United States of America

Faraydon O Karim